hckrnws

Sandboxing is a great security step for agents. Just like using guardrails is a great security step. I can't help but feel like it's all soft defense though. The real danger comes from the agent being able to read 3rd party data, be prompt injected, and then change or exfiltrate sensitive data. A sandbox does not prevent an email-reading agent from reading a malicious email, being prompt injected, and then sending an email to a malicious email address with the contents of your inbox. It does help in implementing network-layer controls though, like apply a policy that says this linux-based sandbox is only allowed to visit [whitelisted] urls. This kind of architectural whitelisting is the only hard defense we have for agents at the moment. Unfortunately it will also hamper their utility if used to the greatest extent possible.

Creator here.

Agreed, sandboxing by itself doesn't solve prompt injection. If the agent can read and send emails, no sandbox can tell a legit send from an exfiltration.

matchlock does have the network-layer controls you mentioned, such as domain whitelisting and secret protection toward designated hosts, so a rogue agent can't just POST your API key to some random endpoints.

The unsafe tool call/HTTP request problem probably needs to be solved at a different layer, possibly through the network interception layer of matchlock or an entirely different software.

We definitely need a vendor-independent tool like this. Have been reviewing the Claude setup and, despite initially being hopeful since it uses bubblewrap, it's quite problematic:

* The definitions of security config in the documentation of settings.json are unclear. Since it's not open source, you can't check the ground truth.

* The built in constructs are insufficient to do fully whitelist based access control (It might be possible with a custom hook).

* Security related issues go unanswered in the repo, and are automatically closed.

Haven't looked into copilot as much but didn't look great either. Seems like the vendors don't have the incentives to do this properly.

So I'm on the lookout for a better way, and matchlock seems like a contender.

There are a lot of options in this space. Armin Ronacher is working on Gondolin (https://github.com/earendil-works/gondolin) for example. I built agentd as a layer in front of this stuff so you can expose secure shell capabilities over the network as a tool rather than baking it into the harness, or running the harness in that environment.

Claude sandbox practically useless IMO. It gives read access to everything by default so its not deny-default.

This is great. Wish this was around when I started working on vibebin ( https://github.com/jgbrwn/vibebin ), probably would have leveraged matchlock instead of Incus/LXC. I guess I could fork/branch and give it a go! Although for vibebin use case I actually need them to not be ephemeral. Edit, ooooh i see `--rm=false` nice

Where do the images come from? What are our options around that and also using custom images etc?

Creator of matchlock here. You can directly use Docker/OCI compatible images (e.g. ubuntu:24.04) as the rootfs with the `--image` flag.

You can also build image with `matchlock build -f Dockerfile -t foo:bar .` - Under the hood it builds the image using buildkit inside the microvm.

Any chance you could look into potentially adding the option to use PVM (eg so a PVM mode instead of KVM) in your matchlock/firecracker implementation?

See https://blog.alexellis.io/how-to-run-firecracker-without-kvm...

Thanks for the response! How would matchlock microvms perform on a KVM VM without CPU passthrough, or is it not possible?

sandboxing is really the only way to make agentic workflows auditable for enterprise risk. we can't underwrite trust in the model's output, but we can underwrite the isolation layer. if you can prove the agent literally cannot access the host network or sensitive volumes regardless of its instructions, that's a much cleaner compliance story than just relying on system prompts.

This may sound obvious, but there must also be an enforcement of what's allowed into that sandbox.

I can envision perfectly secure sandboxes where people put company secrets and communicate them over to "the cloud".

exactly, egress control is the second half of that puzzle. A perfect sandbox is useless for dlp if the agent can just hallucinate your private keys or pii into a response and beam it back to the model provider. it’s basically an exfiltration risk that traditional infra-level security isn't fully built to catch yet.

Sandbox won’t be enough, distroless + “data firewall” + audit

Indeed, but a rock solid sandboxing and isolation strategy is step 0.

This is the confused deputy problem at the application layer. Sandboxing secures the environment, but if the agent has legitimate access to sensitive operations (email, database writes, API calls), prompt injection attacks work through approved channels. The only hard defense is explicit user confirmation for each action, which defeats the point of autonomy.

I've been happily using a container to run my agents [1]. I tried to make it evolve with more advanced features, but it quickly became harder to use and I went back to a basic container which I just start with a run.sh script. Is a similar simple use possible with matchlock?

1:https://github.com/asfaload/agents_container

I use a very similar setup. I initially used nix to manage dev tools, but have since switched to mise and can't recommend it enough https://mise.jdx.dev/

does mise use nix underneath or did you abandon nix entirely?

Mise doesn't use nix. I think the OP is stating he replaced nix with mise.

I think for the first time ever, we are facing a paradigm shift in containment/sandboxing.

Just as Docker became the de facto standard for cloud containerization, we are seeing a lot of solutions attempting to sandbox AI agents. But imo there is a fundamental difference: previously, we sandboxed static processes. Now, we are attempting to sandbox something that potentially has the agency and reasoning capabilities to try and get itself out.

It’s going to be super interesting (and frankly exciting) to see how the security landscape evolves this time around.

I think a sandbox containing a program should only output data. And that data should conform to a schema. The old difference between programs and data instead of turing-complete languages everywhere.

I have been saying for years that technology increasingly requires the development of memetic firewalls - firewalls that don't just filter based on metadata, but filter based on ideas. Our firewalls need to be at least as capable as the entities it seems to keep out (or in).

That sort of firewall is going to be really expensive to run, to the point that it's a financial DOS vulnerability. What is feasible is simpler algorithms that emit alerts on a baseline pattern match, which then get routed to AI observers after some trigger threshold for mitigation. I wouldn't be surprised if someone has already deployed something like that, TBH.

What are the advantages of using this over lxd system container or if we want VM isolation them lxd VMs? Is it the developer experience or there are any agent specific experience which is the key thing here?

The main thing matchlock adds over general-purpose vm/container tooling is agent specific network and filesystem (wip) controls, so if an agent goes rogue it can't exfiltrate your API keys, and damage largely mitigated. You'd have to build all of that yourself on top of LXD (possibly similar to matchlock).

There's also the DX side - OCI image support, highly programmable, fuse for workspace sharing. It runs on both linux and mac with a unified interface, so you get the same/similar experience locally on a Mac as you do on a linux workstation.

Mostly it's built for the purpose of "running `claude --dangerously-skip-permissions` safely" use case rather than being a general hypervisor.

1. Containers aren't a security boundary. Yes they can be used as such, but there is too much overhead (privilege vs unprivileged, figuring out granular capabilities, mount permissions, SELinux/AppArmor/Seccomp, gVisor) and the whole thing is just too brittle.

2. lxd VMs are QEMU-based and very heavy. Great when you need full desktop virtualization, but not for this use case. They also don't work on macOS.

Using Apple virtualization framework (which natively supports lightweight containers) on macOS and a more barebones virtualization stack like Firecracker on Linux is really the sweet spot. You get boot times in milliseconds and the full security of a VM.

qemu has a microvm machine profile, also boots in ms.

There are also tooling on Linux to do containers as microvm's, long before Apple containers were a thing.

And yet Amazon spent a ton of time and money writing Firecracker from scratch for their workloads. Why is that?

Multiple reasons:

1. Firecracker is still a smaller more deliberate surface area 2. qemu didn't have a microvm type at the time. Firecracker was the impetus for it

This is very cool, is it possible to mount NFS as a storage layer?

containers are fine for basic isolation but the attack surface is way bigger than people think. you're still trusting the container runtime, the kernel, and the whole syscall interface. if the agent can call arbitrary syscalls inside the container, you're one kernel bug away from a breakout.

what I'm curious about with matchlock - does it use seccomp-bpf to restrict syscalls, or is it more like a minimal rootfs with carefully chosen binaries? because the landlock LSM stuff is cool but it's mainly for filesystem access control. network access, process spawning, that's where agents get dangerous.

also how do you handle the agent needing to install dependencies at runtime? like if claude decides it needs to pip install something mid-task. do you pre-populate the sandbox or allow package manager access?

Creator of matchlock here. Great questions, here's how matchlock handles these:

The guest-agent (pid-1) spawns commands in a new pid + mount namespace (similar to firecracker jailer but in the inner level for the purpose of macos support). In non-privileged mode it drops SYS_PTRACE, SYS_ADMIN, etes from the bounding set, sets `no_new_privs`, then installs a seccomp-BPF filter that eperms proces vm readv/writev, ptrace kernel load. The microVM is the real isolation boundary — seccomp is defense in depth. That said there is a `--privileged` flag that allows that to be skipped for the purpose of image build using buildkit.

Whether pip install works is entirely up to the OCI image you pick. If it has a package manager and you've allowed network access, go for it. The whole point is making `claude --dangerously-skip-permissions` style usage safe.

Personally I've had agents perform red team type of breakout. From my first hand experience what the agent (opus 4.6 with max thinking) will exploit without cap drops and seccomps is genuinely wild.

Thank you for matchlock! I’ve got Opus 4.6 red teaming it right now. ;)

I think a secure VM is a necessary baseline, and the days of env files with a big bundle of unscoped secrets are a thing of the past, so I like the base features you built in.

I’d love to hear more about the red team breakouts you’ve seen if you have time.

defense in depth makes sense - microVM as the boundary, seccomp as insurance. most docs treat seccomp like it's the whole story which is... optimistic.

the opus 4.6 breakouts you mentioned - was it known vulns or creative syscall abuse? agents are weirdly systematic about edge cases compared to human red teamers. they don't skip the obvious stuff.

--privileged for buildkit tracks - you gotta build the images somewhere.

It tried a lot of things relentlessly, just to name a few:

* Exploit kernel CVEs * Weaponise gcc, crafting malicious kernel modules; forging arbitrary packets to spoof the source address that bypass tcp/ip * Probing metadata service * Hack bpf & io uring * A lot of mount escape attempts, network, vsock scanning and crafting

As a non security researcher it was mind blown to see what it did, which in the hindsight isn't surprising as Opus 4.6 hits 93% solve rate on Cybench - https://cybench.github.io/

that's wild - weaponizing gcc to craft kernel modules is not something I'd expect from automated testing. most fuzzing stops at syscall-level probes but this is full exploit chain development.

the metadata service probing is particularly concerning because that's the classic cloud escape path. if you're running this in aws/gcp and the agent figures out IMDSv1 is reachable, game over. vsock scanning too - that's targeting the host-guest communication channel directly.

93% on cybench is genuinely scary when you think about what it means. it's not just finding known CVEs, it's systematically exploring the attack surface like a skilled pentester would. and unlike humans, it doesn't get tired or skip the boring enumeration steps. did you find it tried timing attacks or side channels at all? or was it mostly direct exploitation?

I'm working on a similar project. Currently managing images with nix, using envoy to proxy all outbound traffic with no direct network access, with optional quota support. Ironically similar to how I'd do things for humans.

My architecture is a little different though, as my agents aren't running in the sandbox, only executing code there remotely.

just from looking at it

on Linux it runs Firecracker: https://github.com/jingkaihe/matchlock/blob/main/pkg/vm/linu...

on macOS uses the Apple's Virtualization.Framework Go wrapper: https://github.com/jingkaihe/matchlock/blob/main/pkg/vm/darw...

nice - I was wondering about the cross-platform story. firecracker on linux for the isolation, virtualization.framework on mac so you don't need vmware.

Why would secrets ever need to be available to the agent directly rather than hidden inside the tool calling framework?

Creator of Matchlock here. Mostly for performance and usability. For interacting with external APIs like GCP or GitHub that generally have huge surface area, it's much more token-efficient and easier to set up if you just give the agent gcloud and gh CLI tools and the secrets to use them (in our case fake ones), compared to wiring up a full-blown MCP server. Plus, agents tend to perform better with CLI tools since they've been heavily RL'd on them.

That doesn't add up to me at all. Agents are RLd on tool usage just as hard and you can provide an "authed API call" tool to whatever you want.

Token efficiency is a good argument actually.

Sometimes people are too lazy to write their own agent loop and decided to run off-the-shelf coding agent (e.g. Claude Code, or Pi in case of clawdbot) in environment.

Exactly.

Is this just a copycat of the deno soundbox announcement from a few days ago?

If I'm already on Linux, how does it compare to using bubblewrap?

Creator here. A few key differences:

1. from isolation pov, Matchlock launch Firecracker microvm with its own kernel, so you get hardware-level isolation rather than bubblewrap's seccomp/namespace approach, therefore a sandbox escape would require a VM breakout.

2. Matchlock intercepts and controls all network traffic by default, with deny-all networking and domain allowlisting. Bubblewrap doesn't provide this, which is how exfiltration attacks like the one recently demonstrated against Claude co-work (https://www.promptarmor.com/resources/claude-cowork-exfiltra...).

3. You can use any Docker/OCI image and even build one, so the dev experience is seamless if you are using docker-container-ish dev workflow.

4. The sandboxes are programmable, as Matchlock exposes a JSON-RPC-based SDK (Go and Python) for launching and controlling VMs programmatically, which gives you finer-grained control for more complex use cases.

See also:

https://github.com/obra/packnplay

https://github.com/strongdm/leash

https://github.com/lynaghk/vibe

(I've been collecting different tools for sandboxing coding agents)

very cool, if you want cross-platform microvms, there's an interesting project called libkrun that powers projects like Podman and Colima.

here's a Go binding: https://github.com/mishushakov/libkrun-go

demo (on Mac): https://x.com/mishushakov/status/2020236380572643720

[dead]

[dead]

Have I told you about our lord and savior: `useradd`

Would you let a pro blackhat loose on your system with just a different user account?

You'd let the pro blackhat loose in your VM on your own system?

No because it's a dumb question and you don't want any stranger inside your home network regardless of firewall.

The comparison you get to make is in terms of the _extra_ security this project buys you.

Might I remind you of two things:

- You're advocating for installing random (?kernel) level software from the internet. That by itself is a real and larger treat than any potentially insecure things my `llm` user _might_ do in the future.

- User accounts security was the goto method for security for a long time. Further isolation was developed to accommodate: 'root' access for tenants, and finer resource limits controls. Neither I care to give an LLM.

So we only have build in firewall and sandbox duplication as the real feature. For the latter, my experience is that it's useless on a personal device, and slows down building or requires too much cache config. I'm not installing random crap, so i can live with the risk of lan exposure.

I'm happy with the maintenance/complexity/threat matrix of useradd.

> You'd let the pro blackhat loose in your VM on your own system?

AWS/GCP/Azure allow that all day every day.

Until you are (or if the agent runs) one privilege escalation away from the whole system being taken over.

So useradd isn't enough.